Actually, the idea of implementing RSI dates back to 2017. This paper is a classic work by Google Brain: *Neural Architecture Search with Reinforcement Learning*. It uses an RNN as the controller to gradually generate the architecture description of sub-networks (layer types, kernel sizes, number of channels, whether to add skip connections, etc.). The generated child network is then trained for real, with validation accuracy used as the reward. REINFORCE is applied to update the controller, enabling it to propose increasingly better architectures. Isn't this essentially using reinforcement learning to train a neural network, then validating its accuracy on the test set, and iteratively continuing the training? https://arxiv.org/pdf/1611.01578
--
24/7 Flashes
More >Today 2026-10-05